Intron, an Africa-focused voice technology platform, has unveiled Sahara v2.5, a new generation of its voice AI platform designed to understand how Africans naturally communicate, including switching between languages within the same sentence or conversation.
Across Africa, people routinely move between languages when speaking. A doctor may explain a diagnosis in English before reassuring a patient in Swahili, while a bank customer may discuss a loan in Yoruba before switching to English or Nigerian Pidgin. Friends, families and colleagues similarly blend languages to convey identity, nuance and context.
However, many AI systems struggle to accurately process such conversations, often missing important sections or requiring users to stick to a single language. Sahara v2.5 is designed to address that gap.
The latest release introduces bilingual speech recognition with language-mixing, also known as code-switching, across 12 African languages, including Zulu, Hausa, Swahili and Luganda. Intron's published benchmarks show Sahara outperforming Gemini, ElevenLabs and Meta across all 12 languages tested.
Intron has also introduced what it describes as the world's first African trilingual speech-recognition model, supporting switching between Kinyarwanda, English and French. The technology uses proprietary algorithms and technology for which the company has filed US patents.
Sahara v2.5 also adds text-to-speech and voice-agent capabilities that support language mixing across 13 languages. According to Intron's benchmarks, the system outperformed ElevenLabs and Gemini in 11 of the 13 languages tested, allowing businesses to build voice experiences, voice bots and voice agents designed to sound more natural to local users.
The platform's speech-recognition coverage has also expanded to include Nupe, Kanuri, Nigerian Fulfulde, Tigrinya, Kikuyu, Dholuo and Somali, taking Sahara's African-language coverage to 31 languages.
At the 2026 Deep Learning Indaba in Lagos, Africa's largest AI gathering, the Sahara CodeSwitch Hackathon brought together more than 120 teams from 21 countries to develop fintech, healthtech, agritech and edutech applications capable of supporting language mixing on Sahara.
Several participating teams compared Sahara with open, closed and commercial models, including Gemini, AssemblyAI and Whisper. The results indicated that Sahara performed significantly better, although it was slightly slower.
The competition continues through September, with a dedicated masterclass hosted by Nvidia on 28 August. Teams are competing for a $10,000 prize pool.
For enterprises, Intron says Sahara v2.5 could reduce the need for customers to change the way they speak to be understood by AI. Contact centres, hospitals, financial institutions, governments and WhatsApp voice assistants can follow conversations as users naturally switch between local languages and English or French.
The company has also deployed offline Sahara models on Nvidia hardware at PAMO Clinics in Port Harcourt, Nigeria, through a global donor-funded project. The deployment is intended to bring private and sovereign AI closer to enterprises, governments and institutions facing regulatory or connectivity constraints that can limit access to advanced AI technology.
Sahara v2.5 introduces streaming speech recognition and text-to-speech capabilities through its API, supporting live captions, real-time applications and real-time speech generation.
Intron said improvements in latency, concurrency and reliability are designed to support high-volume applications, including chatbots, contact centres, medical documentation, financial services, legal workflows, and agricultural and climate advisory services.
Sahara already supports organisations across healthcare, legal services, financial services and contact centres. Existing deployments include:
Branch International — Financial services
Sahara-powered Branch collections agents recovered more than ₦1.2 million in delinquent loans in one week, with record after-hours and weekend repayments. The agents outperformed human agents in recovering delinquent loans that had been outstanding for more than 356 days.
Audere Africa — Voice chat
Young people across South Africa can now speak naturally with Audere's Self-Cav reproductive health WhatsApp chatbot using voice, rather than being restricted to text conversations.
During field testing with South African English accents, Audere found Zulu code-switching to be the leading source of transcription errors. The organisation welcomed Intron's introduction of code-switching support for Zulu and Afrikaans.
Ogun State Judiciary — Legal
The Ogun State Judiciary has worked with Intron for more than a year, expanding from an initial pilot court to nine courts, with plans to automate transcription across all 18 high courts in the state.
Court proceedings that previously took more than four hours can now be completed in half the time, allowing judges to follow courtroom dialogue without repeatedly stopping to write notes.
Healthcare
A 14-minute Swahili-English doctor-patient consultation in Nairobi can be converted into a structured clinical note in less than 30 seconds, reducing the documentation burden on physicians and allowing them to focus more on patient care.
Adanne Anene, Head of Product Africa at Branch International, said: “Collaborating with Intron to build a Branch-aligned collections bot was a rewarding experience; customers engaged naturally even after hours and on weekends. Conversations were human and effective at delivering real payments on delinquent loans, sometimes over two years old. Adding local language and language-mixing support in this v2.5 release helps us reach even more customers.”
Sahara currently supports production and research deployments for more than 40 enterprise customers across six countries: Nigeria, Kenya, South Africa, Uganda, Rwanda and Ghana.
Intron has published benchmark results comparing Sahara with global models on African code-switched speech. Measured using word error rate (WER), where a lower score indicates fewer errors, Sahara v2.5 recorded an average WER of 34.3% across 12 languages, compared with 53.8% for Gemini 3.6.
The results represent a 36% relative reduction in errors, or an average difference of 19.5 percentage points per language, according to Intron, demonstrating a substantial and consistent advantage over global models on African code-switched speech.
Separate benchmarking conducted through Gooey.ai for the Gates Foundation and CLEAR Global found Sahara performed better on five of the seven Nigerian languages evaluated, including in comparisons with Gemini and Meta's Omnilingual model.
“Code-switching was one of the biggest problems that consistently came up for clients deploying real-world voice AI. The content loss with most models increases post-editing time, and propagates errors and omissions to downstream summaries, call logs, court records, and clinical notes,” said Tobi Olatunji, CEO of Intron. "Africa needs AI built for how Africans really speak. People should not have to translate themselves for a machine, flatten their accent, avoid local expressions or repeat only the English part of what they said. Voice AI should work with the way people already speak."
Alongside the launch of Sahara v2.5, Intron has published its 2026 Africa Voice AI Report. The report challenges the assumption that collecting African-language data is the only major barrier to reliable voice AI, arguing that research capacity, orchestration and implementation expertise are equally important.
The report highlights ambient medical scribes as an example. Although the technology has been widely adopted in the US and Europe, it has struggled in some African clinics where consultations frequently move between multiple languages.
Since raising $1.6 million in pre-seed funding in 2024, Intron has expanded its training data to more than 150,000 hours of African-language audio from over 53,000 speakers. The dataset covers 64 languages and more than 500 accents.
Sahara's current code-switching support includes Afrikaans-English, Akan-English, Amharic-English, Hausa-English, Igbo-English, Kinyarwanda-English, Kinyarwanda-French, Luganda-English, Nigerian Pidgin English, Swahili-English, Wolof-French, Yoruba-English and Zulu-English.
No comments:
Post a Comment